ECMWF/ H-SAF and HEPEX Workshops on coupled hydrology
Reading, 3-6 November 2014
Assimilation of satellite soil moisture data in a
distributed hydrological model: impact on the
hydrological cycle in some Italian basins
S. Gabellani and P. Laiolo
[email protected]
CIMA Research Foundation
International Centre on Environmental Monitoring
Data Assimilation
Data assimilation is used operationally in
oceanography and meteorology, but in
hydrology it is only recently that international
research activities have been
deployed.
(some) Open questions in DA
1. Which is the best DA techniques?
2. How can satellite data be used in a framework for DA
in hydrological models?
3. Which is the proper model configuration?
4. Which is the impact of DA on the hydrological cycle?
1. Data Assimilation Technique
Direct insertion (Houser et al. 1998; Walker et al. 2001a)
Statistical correction (Houser et al. 1998)
Successive correction Bergthorsson and Döös (1955)
Analysis correction Lorenc et al. (1991)
Nudging (Stauffer and Seaman 1990)
Optimal interpolation (Lorenc et al. 1991)
Kalman Filters, simple, extended, ensemble (Evensen)
Particle filter (Kalman, 1960; Evensen 1994, Gordon et al. 1993)
3D & 4D var -> Var. filter
SEQUENTIAL
FILTERS
xki+ = xki- + Gk(yki - xki- )
Houser, De Lannoy and Walker (2012). Hydrologic Data Assimilation, Approaches to Managing
Disaster - Assessing Hazards, Emergencies and Disaster Impacts,
http://www.intechopen.com/books/approaches-to-managing-disaster-assessing-hazardsemergencies-and- disaster-impacts/land-surface-data-assimilation
1. Data Assimilation Technique
The assimilation technique is particularly important in some
cases
Samuel, J. et al. 2014 (JoH)
“[…] In the streamflow assimilation, soil moisture states were markedly Distorted […]”
”General filtering approaches in hydrologic data assimilation, such as the ensemble
Kalman filter (EnKF), are based on the assumption that uncertainty of the current
background prediction can be reduced by correcting errors in the state variables
at the same time step. However, this assumption may not be valid when assimilating
stream discharge into hydrological models to correct soil moisture storage due to the
time lag between the soil moisture and the discharge …”
Li et al. 2013 (WRR)
The EnKF is designed to update model-forecasted state predictions at the same time
an observation is acquired. No attempt is made to reanalyze previous model
predictions in response to a particular observation. In contrast, the Ensemble Kalman
Smoother (EnKS) can be used to update all model states predictions within a fixed lag
of past time (Dunne and Entekhabi, 2005).
Crow and Ryu, 2009 (HESS)
2. How can sat. data be used in DA?
Satellite data give information of soil moisture for the
first centimetres of the soil. This may not match the
layer depth simulated by the model (different
climatology and considerable bias)
root-zone
Usually satellite
soil moisture data
CANNOT be
directly used within
hydrological
models
2. How can sat. SM data be used in DA?
A. “Transform” the sat. SSM in the “same” modelled
variable
•  Filtering
B. Adjusting the observation to match the climatology of
the model
•  Bias handling
2. How can sat. SM data be used in a DA?
Filtering: A filtering technique is applied to obtain information of a deeper
soil layer
Wagner et al., 1999, Stroud, 1999
Albergel et al., 2008
SWI:
Soil Water Index
t: time
SSMti: relative Surface Soil
Moisture [0,1]
ti: acquisition time of SSMti
T: characteristic time length
SSM
SWI
2. How can sat. SM data be used in DA?
Bias Handling: Several potential strategies exist and have been
applied in hydrologic data assimilation
Variance matching (VM) (Brocca et al. 2010, 2012, Matgen et al. 2011, Chen et al. 2011)
Linear regression techniques (LR)
Cumulative distribution function matching (CDF) (Reichle and Koster 2004)
Anomaly based cumulative distribution (aCDF)
Triple collocation analysis-based approach (TCA) (Stoffelen 1998, Yilamz and
Crow 2013)
There many methods their optimality (for real cases) in terms of
error analysis in an assimilation framework has not been yet
analysed
2. How can sat. SM data be used in DA?
1. Filtering -> SWI
2. Bias handling
SAT * =
SAT * =
SAT - m ( SAT )
s ( SAT )
× s ( SDmod ) + m ( SDmod )
SAT - min ( SAT )
éëmax ( SAT ) - min ( SAT )ùû
×éëmax ( SDmod ) - min ( SDmod )ùû + min ( SDmod )
2. How can sat. SM data be used in DA?
3-D EnKF
Sahoo et al., 2013
After the assimilation the analysis is
bias-corrected to bring the output to the
true climatology
The 1-D EnKF application assimilates a
priori partitioned observations at the fine
scale model grid cells.
The 3-D EnKF algorithm downscales the
coarse
observations
within
the
assimilation scheme and uses multiple
coarse observation grid cells, as shown
in Fig. 2.
Both the EnKF algorithms produce finescale results that are closer to the in
situ data than either the model open
loop or the satellite observations alone.
The 3-D EnKF slightly outperforms the
1-D EnKF and better preserves realistic
spatial patterns because of the colored
spatial error correlations and the
corresponding impact of multiple coarse
observation grid cells
3. Proper model configuration
Filtering the observation
Modifying the model
structure
Chen et al. 2011 (AWR)
Brocca et al. 2012 (IEEE ToGRS)
Flores et al. 2012 (WRR)
4. Which is the impact of DA on the hydrological cycle?
Many of the hydrologic DA studies reported in the literature focused on advancing the
theoretical development of DA techniques using synthetic experiments (e.g.,
Andreadis et al., 2007; Kumar et al., 2009; Crow and Ryu, 2009).
• diagnostic and design purposes such as assessing the impact of improper
characterization of model and observation errors (e.g., Crow and Van Loon, 2006;
Reichle et al., 2008
• evaluating the potential benefits of future satellite missions (e.g., Matgen et al.,
2010)
Only a few formulated DA in an operational setting and attempted to evaluate the
performance gain from DA in real cases (e.g., as a result of better characterized initial
conditions) studies (e.g., Seo et al., 2003, 2009; Thirel et al., 2010; Weerts et al., 2010;
DeChant and Moradkhani, 2011, Brocca et al. 2012 )
“There
is a strong need to estimate soil moisture content through assimilating
remotely sensed soil moisture into a long-term, physically based distributed
catchment scale hydrologic model. Most of the previous studies that explored
DA for runoff simulation used conceptual rainfall-runoff models (Aubert et al.
2003; Weerts and El Serafy, 2006; Crow and Ryu, 2009; van Delft et al. 2009) or
lumped models (Jacobs et al 2003) or for short-term period with real
measurements (Pauwels et al. 2001)”. Han et al. 2012
4. Which is the impact of DA on the hydrological cycle?
Chen et al. 2011 (AWR)
4. Which is the impact of DA on the hydrological cycle?
Han et al., 2012
Synthetic experiments using SWAT
model Results of assimilation:
• great impact on soil moisture
• small impact on discharge
• impact on discharge is a function of soil
type
• the capability of the SSM assim. for
improving streamflow is constrained by
the accuracy of precipitation
Assimilation of sat. SM in distributed hydrological model
MAIN CHARACTERISTICS:
• Simple but complete description of Hydrological Cycle
•
•
Schematization of vegetation interception and water table
Tank schematization of overland and channel flows
• Mass Balance and Energy Balance completely solved
Silvestro et al., 2013
• Fully Distributed
• River network derived from a DEM
• Spatial-temporal evolution of:
•
•
•
•
•
•
Streamflow
Evapotranspiration
Vegetation retention
Land Surface Temperature
Soil Moisture
Water table
The model Fortran code is open and can be requested
to: http://www.cimafoundation.org/cimafoundation/continuum/
• It can be calibrated using only satellite data (e.g. surface
temperature or soil moisture). Model suitable for application in data
scarce environments
Silvestro et al., 2014
Continuum
Saturation Degree of root zone
0 £ SD £1
V(t)= Actual water volume
Vmax= Max soil retention capacity (related to soil
type and land use through the CN)
root-zone
Time frequency: Hourly map
Resolution: 100 m
Italian test basins
MAGRA river
1700 km2
Calamazza
ORBA river
CASENTINO river
Casalcermelli
880 km2
800 km2
Subbiano
H-SAF Soil Moisture products
• SM-OBS-1
(H07)
Large-scale surface soil moisture (SSM) [-]
Time frequency: 2 maps per day, 1-2 days revisit time
Spatial coverage: Strips of 1000 km swath covering the whole globe
Resolution: 25 km
• SM-OBS-2 (H08)
Small-scale surface soil moisture (SSM) [-]
Time frequency: 2 maps per day, 1-2 days revisit time
Spatial coverage: Strips of 1000 km swath crossing the H-SAF area
Resolution: 1 km
• SM-DAS-2 (H14)
Profile Soil Moisture Index (SMI) in the roots region [-]
Time frequency: Daily map (at 00.00)
Spatial coverage: Globe
Horizontal resolution: 25 km
Vertical resolution: 4 layers (0-7 cm, 7-28 cm, 28-100 cm, 100-289 cm)
SMOS soil moisture product
● Level 2 Soil Moisture
volumetric soil moisture content (SMC) [m3/m3]
Time frequency: 2 maps per day, max 3 days revisit time
Spatial coverage: 600 km swath covering the whole globe
Resolution: 43 km in average, 35 km (centre of field of view)
Data Preparation
• Satellite soil moisture data regridded to Continuum
SSM
SMC
SMI
(H07 – H08)
(SMOS)
(H14, mean L1,L2)
Normalization
Normalization
grid using nearest neighbour method
• Assimilation of the mornig passes only
• Discarded H07 data with high quality flag
• Discarded SMOS data with DQX>=0.045 and
Exp. filter
Exp. filter
SWI
SWI
Min Max Corr.
Min Max Corr.
Linear Resc.
SWI*
SWI*
H14*
Assimilation
Assimilation
Assimilation
RFI/200>1
Normalization:
Linear Rescaling (H14)
Min Max correction (H07, H08 and SMOS)
Assimilation in Continuum model
• Assimilation of the four SSM products
1. Model scale  Re-grid of Sat. data at model’s
resolution, filtering and bias handling and
Assimilation
2. Sat. scale  Re-grid of Model’s state at Sat.
resolution, filtering and bias handling,
Assimilation and then downscaling of
assimilated state to model’s resolution
3. Model scale  Re-grid of Sat. data at model’s
resolution and Assimilation
Nudging
Ensemble
1. Model Scale  Re-grid of Sat. data at model’s
resolution and Assimilation
H07
Regrid on
DEM grid
(100m)
Data flag
+ Exp.
Filter
MinMax
correction
Example of the scheme
used for the assimilation of
H07 product
Assimilation in
hydrological model
2. Sat. Scale  Re-grid of Model’s state at Sat.
resolution, Assim. and then downscaling to model’s
resolution
H07
Regrid on
DEM grid
(100m)
H07 Regrid
on reg. grid
Data flag
+ Exp.
Filter
Example of the scheme
used for the assimilation of
H07 product
MinMax
correction
Assimilation in
hydrological model
at coarse resolution
1. Model Scale  Re-grid of Sat. data at model’s
resolution and Assimilation
Nudging
assimilation
scheme
X+mod= New Saturation Degree
X-mod = Background modeled Saturation Degree
Xobs= Observed Saturation Degree
SWI* (H07, H08, SMOS)
SMI* (H14)
No assimilation over
urban areas and rivers
G = Gain
RMSDmod = Root Mean Square Difference of X-mod = 0.092
(Estimated from a study over modeled soil moisture outputs)
RMSDH14: 0.22 [-]
RMSDobs= Root Mean Square Difference of Xobs
(SOURCE: Albergel validation work presented during H-SAF meeting in Budapest 2013)
RMSDSWI.HSAF: 0.12 [-] for H07 and H08
(SOURCE: Brocca et al. 2011)
RMSDSWI.SMOS: 0.24 [-]
(SOURCE: Albergel et al. 2012)
Nudging assimilation scheme
Satellite scale
X+mod= New Saturation Degree
X-mod = Background modeled Saturation Degree
Xobs= Observed Saturation Degree
G = Gain value
SWI* (H07. H08. SMOS)
SMI* (H14)
G = 0.3 (H14)
G =0.43 (H07 and H08)
G = 0.28 (SMOS)
H = Observation operator (allow to obtain the map at 12.5 km resolution from that at 100 m resolution)
R = Regrid operator (allow to obtain the map at 100 m resolution from that at 12.5 km resolution)
S = Spatialization operator (allow to redistribute the correction on the 100 m grid. The correction depends on
the ratio between the value of X-mod at each 100 m pixel and the mean soil moisture value at the
corresponding 12.5 km pixel)
Bayesian assimilation scheme
Model scale
SDass= Posterior mean of Saturation Degree
SDmod(t) = Modeled Saturation Degree
SWI
SDobs(t) = Observed Saturation Degree
SMI (H-14)
R = Variance of SDoss = 0.04 (assumption)
m = Expected value of SDmod
P = variance of SDmod
N = 20 parameters sets
Soil moisture basin scale comparison
Orba
Period: July 2012 – June 2013
R= 0.82
R= 0.85
R= 0.97
R= 0.84
Soil moisture basin scale comparison
Casentino
Period: July 2012 – June 2013
R= 0.89
R= 0.86
R= 0.86
R= 0.68
Soil moisture basin scale comparison
Magra
Period: July 2012 – June 2013
R= 0.66
R= 0.45
R= 0.68
R= 0.12
Annual results - Orba
EOL = 0.63
MAE = 17.4 [m3/s]
RMSE = 25.3 [m3/s]
1 n
MAE   Qsi  Qoi
n i 1
n
E  1
RMSE 
 Qo  Qs 
i 1
n
2
i
i
2


Qo

Qo
 i
i 1
1 n
Qsi  Qoi 2

n i 1
Qsi – simulated values
Qoi – observed value
Annual results - Casentino
EOL = 0.70
MAE = 14.3 [m3/s]
RMSE = 21.6 [m3/s]
Annual results - Magra
EOL = 0.72
MAE = 28.4 [m3/s]
RMSE = 46.7 [m3/s]\
Seasonal results - Orba
Model Scale - Nudging
Sat. Scale - Nudging
ORBA - E Improvements respect OL
Nudging assimilation - Model scale
100
89
ORBA - E Improvements respect OL
Nudging assimilation - Satellite scale
95
100
Assim H07
Assim H07
80
80
Assim H08
60
48
22
5
0
Assim SMOS
15
20
15
13
11
6
4
0
-80
-80
-100
-100
g
-60
r in
Sp
n
-24
-40
-60
-2
-6
r
um
ut
-20 -24
te
in
W
A
er
m
-20
m
Su
-15
g
-25
-5
r in
Sp
-40
-15
-4
r
n
er
m
um
ut
7
te
in
W
A
m
Su
-20
6
2
0
20
[%]
[%]
Assim SMOS
20
Assim H14
35
34
40
19
20
Assim H08
60
Assim H14
40
98 95
ORBA - E Improvements respect OL
Bayesian assimilation
100
92
Assim H07
80
69
Assim H08
52
60
40
61
37
43
Assim H14
43
Assim SMOS
21
[%]
20
n
g
-18 -14
r in
Sp
r
um
ut
er
m
te
in
W
A
m
Su
-20
1
-17
0
-40
-60
-80
-54
Model Scale - Ensemble
-100
-58
-68
-85
Summer
Autumn
Winter
-2.64
0.57
0.52
Spring
0.78
Seasonal results - Casentino
CASENTINO - E Improvements respect OL
Nudging assimilation - Model scale
100
93
Assim H07
80
60
43
47
40
25
[%]
20
10
Assim H08
Summer
Autumn
Winter
Assim H14
-1.50
0.50
0.86
Spring
-0.64
Assim SMOS
12
3
1
0
0
0
-6
r in
Sp
n
g
r
um
ut
er
m
0
0
te
in
W
-2
A
m
Su
-20
75
-25
-40
-60
-80
Model Scale - Nudging
CASENTINO - E Improvements respect OL
Nudging assimilation - Satellite scale
-100
100
Assim H07
80
65
60
47
40
27
[%]
20
Assim H08
Assim H14
30
22
20
21
3
Assim SMOS
4
-5 0
0
r in
Sp
-1
g
r
n
er
m
um
ut
-5
te
in
W
A
m
Su
-20
14
-40
-43
-60
-80
-100
Sat. Scale - Nudging
-93
Seasonal results - Magra
MAGRA - E Improvements respect OL
Nudging assimilation - Model scale
240
Assim H07
200
193
174
Assim H08
160
Assim H14
[%]
120
Summer
Autumn
Winter
-0.18
0.84
0.60
Assim SMOS
80
Spring
0.24
61
40
3
0
1
3
-2
0
0
r in
Sp
r
g
te
in
W
-4
um
ut
n
-68
MAGRA - E Improvements respect OL
Nudging assimilation - Satellite scale
-102
-116
231
240
Model Scale - Nudging
207
Assim H07
200
Assim H08
160
Assim H14
120
[%]
-120
er
m
-80
0
A
m
Su
-40
-5
2
124
Assim SMOS
80
40
3
0
7
8
0
r in
Sp
r
Sat. Scale - Nudging
g
te
in
W
n
-71
um
ut
-44
0
-5
A
-120
-37
-20
er
m
-80
0
-3
m
Su
-40
1
Discharge events results - Orba
Nudging – Model scale
Discharge events results - Orba
Efficiency - H14 assimilation
Efficiency - H07 assimilation
1,0
1,0
0,8
0,8
0,6
0,6
0,4
0,4
0,2
0,2
0,0
0,0
E
E
-1,6
-1,2
Nud Sat
-1,4
Bayes Mod
-1,6
Efficiency - H08 assimilation
Efficiency - SMOS assimilation
1,0
1,0
0,8
0,8
0,6
0,6
0,4
0,4
0,2
0,2
0,0
E
-0,4
OL
Nud Mod
-1,2
Nud Sat
-1,4
Bayes Mod
-0,6
-0,8
-1,0
-1,2
-1,4
-1,6
OL
Nud Mod
Nud Sat
Bayes Mod
.8
Ev
.7
Ev
.6
Ev
.5
Ev
.4
Ev
.3
Ev
.2
Ev
-0,2
.1
Ev
.8
Ev
.7
Ev
.6
Ev
.5
Ev
.4
Ev
.3
Ev
.2
Ev
.1
Ev
E
0,0
-0,4
-1,6
.8
Ev
Bayes Mod
-1,0
.7
Ev
-1,4
-0,8
.6
Ev
Nud Sat
-0,6
.5
Ev
Nud Mod
-1,0
-1,2
-0,2
.4
Ev
OL
-0,8
Nud Mod
-1,0
-0,4
-0,6
OL
-0,8
.3
Ev
-0,6
-0,2
.2
Ev
-0,4
.1
Ev
.8
Ev
.7
Ev
.6
Ev
.5
Ev
.4
Ev
.3
Ev
.2
Ev
.1
Ev
-0,2
Impact of assimilation on
other state variables
• Water Volume (V)
• Evapotranspiration (Evt)
• Land Surface Temperature (LST)
Model calibration with satellite data
Parameter calibration using SWI(H07)
Satellite data reduced
hydrological
uncertainty and could
be used to calibrate
models
Val. Period: 1/06/2009 – 31/12/2011
Calibration
results using
only
geomorphology
(DEM) and
SWI from H07
Nash and
Sutcliffe’s
efficiency
coefficient
NSDisch
0.81
NSSWI
0.79
Conclusions
• Annual evaluation
– Assimilations of Soil moisture products improved the performances
– “Sat. Scale” is better than “Model Scale” for Magra and Orba
– The Ensemble method is promising on Orba
• Seasonal evaluation
– Summer and Autumn benefit most from assimilation
– “Sat. Scale” is better than “Model Scale” for Magra and Orba
• Events evaluation
– H14 leads to improvement in 90% of cases
– H07 and H08 lead improvement in 50% of cases
– SMOS lead improvement in 35% of cases
Thanks to L. Campo, F.
Silvestro, F. Delogu, R.
Rudari, L. Pulvirenti, G. Boni,
N. Pierdicca, L. Brocca, C.
Massari, L. Ciabatta, S.
Hasenauer, S. Puca

201 Dickens Road Fuquay-Varina, North Carolina

35 years of remotely sensed soil moisture – ECV SM 02.0